A novel approach to the problem of computing the direction of arrival (DOA) of a sound source using a two-element microphone array is presented. Typically, the DOA is computed by peak picking from the resulting cross-correlation function. In order to improve such estimates it is usually desirable to pre-filter the signals prior to cross correlation. However, proper selection of these pre-filters is often problematic. The proposed algorithm adapts the filters during an initial training session to directly minimize the number of location estimation errors. The only information provided is the signal DOA, no explicit knowledge of the signal or noise spectra are required. It is assumed however that the overall signal/noise statistics are long term stationary over the training and subsequent testing. Examples of both linear and nonlinear filter design are presented for a talker location identification task.